AI with no restrictions: connect in two minutes

Connect to an AI with no restrictions in two minutes

Three settings. Create the key in the console, point your client at the base URL, name a model id from the model list. The client can be the OpenAI SDK, SillyTavern, Janitor AI, Open WebUI or a curl line.

Base URL

https://api.uncensoredllm.ai/v1

Key

sk-ullm-live-…

Created in the console, shown once

Model

glm-5.3-uncensored

Any id from the model list

The Python version, streaming, with a question most hosted assistants decline:

The same request as curl, printing the cost header the gateway returns with every answer:

x-uplink-charge carries the settled charge for that request and x-uplink-request-id identifies it for support. A request that fails before the model answers is not charged. The balance is prepaid from $10, settles per token and never expires.

What it costs per token

Ten of the listed builds, read from the catalog when this page rendered. The full list, sortable by price and context, is on the Models page; the top-up tiers and the per-token rules are on Pricing.

Top up from $10. Bonus credit by top-up size: +5% from $200, +10% from $500, +15% from $1,000. No subscription and no monthly minimum; a balance that sits unused is still there next year.

Worked example at today's rates: one chat turn of 1,500 tokens in and 500 tokens out on glm-5.3-uncensored settles at $0.0009, and the same turn on glm-5.3-flash-uncensored at $0.0004. Both figures are computed from the catalog rates above, so they move when the rates move.

What "no restrictions" means in the weights

A chat model refuses because its training taught it to. Arditi and colleagues showed in 2024 that this refusal behaviour in open-weight chat models is carried by a single direction in the model's internal activations. Project that direction out of the weights and the model stops refusing, while its other abilities stay close to where they were. The paper is Refusal in Language Models Is Mediated by a Single Direction. The community name for the edit is abliteration.

An abliterated model carries no instruction to comply. The part of it that said no is gone. That is the whole difference from a jailbroken AI: a jailbreak is a prompt that talks a model around its training, it rides on every turn as paid context, it stops working when the model updates, and the model can still drop back into refusal forty messages into a long chat. With abliteration the change is in the weights, so there is nothing to maintain and nothing to break.

An uncensored fine-tune reaches the same place by a different road. The model is trained further on conversations that contain no refusals, which is how the Dolphin builds and most roleplay tunes are made. Abliterated vs uncensored measures the two methods on this catalog's builds with the daily check.

Three places a restriction can live

A restriction sits in one of three layers, and only the first is the model. The weights carry the trained refusal; abliteration and uncensored fine-tuning work here. A system prompt the vendor prepends can tell an otherwise willing model what to decline; you never see it in a hosted app, and by API you write it yourself. A moderation layer in front of the model can block the request before it is sent or cut the answer after it comes back; it works on any model, trained to refuse or not.

A product that is unrestricted in the second and third layers and restricted in the first still refuses, which is what the hidden-filter complaint below describes. A model that is unrestricted in the first layer, reached through an API that adds nothing in the other two, has no layer left to refuse from.

The hidden-filter objection

One result for this query argues that "no limits" chat apps are ordinary interfaces with a filter hidden behind them. The complaint is fair about an app that puts a moderation layer in front of a model trained to refuse: the app can relax the layer, and the model still says no. It does not apply to a model whose refusal was removed in the weights and is served through a plain API. The gateway behind this site adds no moderation layer to chat requests; the request reaches the model you named and the answer comes back as the model wrote it. Image and video prompts carry two guards: nothing that depicts a minor (the acceptable use page says why), and no sexual edit of a photo of a real person.

Whether the model you named is the model answering is a separate question, and it has a daily answer. The check at /proof sends every model a fixed prompt each day and records a verdict per model. 125 of the 133 listed chat models that were checked came back stable on October 24, 2026.

Why it matters

In a roleplay client a refusal on turn 40 breaks the character and the scene. The writer edits the message, resends it and pays for the turn twice. On a model with the refusal removed the turn goes through the first time, and the system prompt can describe the character instead of arguing with the model.

In an agent loop a refusal is a dead end. The loop either stops or retries, and a retry against a model trained to refuse returns the same answer at the same price. A security team that asks a model to read an exploit, classify a phishing kit or rewrite a payload for a detection rule needs a model that reads the input instead of lecturing about it. The security builds in the catalog exist for that job.

The money side is plain. A jailbreak prompt is paid context on every turn, so its cost scales with the length of the chat. A consumer app with no restrictions is free with a daily cap or priced per month, whether you send ten messages or ten thousand. By API you pay the per-token rate in the table above and nothing else, and the bill for a quiet month is zero.

The last reason is control. By API you choose the model, the client, the context length and the system prompt. In a hosted chat app the vendor makes those four choices for you and can change any of them tomorrow.

Types of unrestricted AI, with catalog ids

"Uncensored" covers five kinds of build. They differ in how the refusal was removed, what they are good at and what they cost. Every id below is live in the catalog today and has its own page with the price, the context window and the daily check verdict.

Abliterated builds

The refusal direction is removed from a released model's weights; nothing is added. The base model's knowledge and style stay. Examples: glm-5.3-abliterated, glm-5.3-flash-apex-abliterated, qwen3.8-27b-obliterated, llama-3.3-70b-abliterated and gemma-4-31b-cognitive-unshackled. Pick one when you want the behaviour of a known base model without the refusals. How the edit changes answers, measured on these builds, is in abliterated vs uncensored.

Uncensored fine-tunes

A base model trained further on refusal-free data. These are the general assistants of the catalog, listed first on the Models page: glm-5.3-uncensored, deepseek-v4-pro-uncensored, kimi-k3-uncensored, minimax-m3-uncensored, dolphin-mistral-24b-venice and hermes-4-405b. Pick one for coding, research, long-form writing or an agent. The full list sorts them by price.

Roleplay tunes

Fine-tunes built for character work and fiction, usually by one named author on a 12B to 70B base. They hold a persona, write in scene and follow a character card: cydonia-24b-v4.3, magidonia-24b-v4.3, roleplay-uncensored, qwen3.8-27b-novelist and euryale-70b-v2.3. They are the builds SillyTavern and Janitor AI are set up for, and no-filter AI chat by API goes through them one by one. The SillyTavern fields are in the SillyTavern guide.

Security builds

Models tuned on security corpora with the refusal removed, for red teams, detection engineering and malware analysis: glm-5.3-flash-cybersecurity and cyberheabsy. They read exploits, logs and payloads as text to analyse. Context on the flash build is 1,048,576 tokens, enough for a packet capture or a decompiled binary in one request.

Uncensored image and video models

The same idea applied to media: diffusion and video models served with the safety checker off, priced per image or per second of video. The media catalog lists them with every price tier and the request format.

Which type to pick

• You know and like a base model and only want the refusals gone: its abliterated build. The style, the knowledge and the context window are the base model's.
• You want the strongest general assistant for code, research or an agent: an uncensored fine-tune of a current large open model. Start with glm-5.3-uncensored and move up the price table if the task needs it.
• You write fiction or run character chats: a roleplay tune, chosen by context window first and by the author's style second. The Gemma 4 and Qwen 3.8 families each have a dozen or more roleplay siblings on the Models page, most within a few cents of each other per million tokens, so trying several costs little more than staying on one.
• You analyse malware, phishing kits or exploit code: a security build. Both carry 1,048,576 tokens of context; the price table on the Models page separates them.
• You need pictures or clips: a media model, by the per-image or per-second price in the table and the endpoint your code already calls.

Chat apps, the zero-setup option

Venice, FreedomGPT, Shannon AI, NoLimitGPT and aiwithnorestrictions.com rank for this query as hosted chat apps (Google, read 2026-10-24). They need no client and, in the case of aiwithnorestrictions.com and FreedomGPT, no account; Venice's result promises private, uncensored chat with no credit card. For a quick conversation in a browser tab they are the shortest path. The vendor chooses the model and whatever sits in front of it, and the price is per month. By API you make those choices and pay per token.

How to set it up, step by step

1. Create a key in the console. It is shown once; store it in your client or a secrets manager, never in a repo.
2. Top up. The minimum is $10; bonus credit applies by size (+5% from $200, +10% from $500, +15% from $1,000). The balance settles per token and does not expire.
3. Pick the model by job. A general assistant: glm-5.3-uncensored. Long roleplay: cydonia-24b-v4.3, with 131,072 tokens of context. Reasoning-heavy work: deepseek-v4-pro-uncensored. Drafts at the lowest rate in the table: glm-5.3-flash-uncensored. Each id's page shows the price, the context window and today's check verdict.
4. Put the three settings into the client. The field names differ by client and are listed in the tools section below; the values are always the base URL, the key and the model id.
5. Send one test request and read x-uplink-charge on the answer. That number is what the balance moved by. Streaming answers end with the usage block, so a client that counts tokens sees the same figures.
6. Set a spend limit on the key in the console before you paste it into a shared tool. Usage per key is on the same page.
7. Open /proof and find your model. A stable verdict means the fixed daily prompt came back from the model you named. If a build ever shows changed, its row says since when.

Best practices

• Drop the jailbreak block. Keep the system prompt to the persona, the format and the facts the model needs. An abliterated model follows instructions; it does not need to be argued with, and every token of argument is billed on every turn.
• Match the context window to the chat. mythomax-l2-13b has 4,096 tokens; cydonia-24b-v4.3 has 131,072 tokens; qwen3.8-27b-obliterated has 524,288 tokens. A 200-message scene needs the second or third, and the client's context limit should sit under the model's.
• Draft on the flash build, finish on the big one. The price table shows the gap. Run glm-5.3-flash-uncensored for exploration and switch to glm-5.3-uncensored or deepseek-v4-pro-uncensored for the pass that ships.
• Stream, and set max_tokens. Streaming shows the first words while the rest is still being written and lets you stop a turn that went wrong before it runs to the limit. A max_tokens that fits the turn keeps a runaway answer cheap.
• Use the OpenAI chat-completions format. Every listed chat model answers it, so the OpenAI SDK in any language works with the base URL swapped. The docs page has the Python, Node and Anthropic-SDK versions.
• Read the model page before you switch. Price, context, wire formats and the daily verdict are on one screen, so a change of model is a change of one string in the client and nothing else.

Common mistakes

• Pasting a jailbreak prompt in front of an abliterated model. It adds paid tokens to every turn, and several roleplay tunes will answer the jailbreak's theatrics in character instead of getting on with the scene.
• Expecting knowledge the base model never had. Abliteration removes a behaviour. It adds no training data, so a 13B tune from 2023 still knows what a 13B tune from 2023 knew.
• Judging "no restrictions" by the app. The filter can live in the model, in a layer in front of it, or both. Ask which model the app runs and whether it is abliterated; if the vendor will not say, assume a trained-to-refuse base behind a relaxed filter.
• A short-context tune for a long chat. When the context fills, the client silently drops the oldest messages and the character forgets the scene. Check the context column before a long session.
• Calling the wrong wire format. Every listed chat model answers the OpenAI format at /chat/completions. A model page lists any other format it also answers; if the Anthropic Messages format is not listed, the request belongs on the OpenAI route.
• A key with no spend limit in a shared tool. A community bot or a forwarded proxy URL can run a balance down in an afternoon. Set the limit per key, and give each tool its own key.
• Choosing by the word "uncensored" alone. It covers an 8B roleplay tune and a 405B assistant. The price table and the context column are the real choice; the word is only the filter that got you to the list.

Tools: which client for which job

Every client below takes the same three values. The table says where each one puts them.

SillyTavern's own documentation on API connections explains the Chat Completion source list and the proxy settings in detail; the custom endpoint field it describes is the base URL above.

Questions